ScaleSense: an improved method of automatic identification of ruler scales on images to solve the problem of estimating the geometric dimensions of certain objects ; ScaleSense: удосконалений метод автоматичного визначення масштабу лінійки на зображеннях для вирішення завдання оцінювання геометричних розмірів певних об'єктів

The identification of pixel correspondence to real-world units in an image has been identified as a crucial step for further analysis of the geometric characteristics of objects in the image. An enhanced iterative method for the automatic determination of ruler scale on the image based on digit recognition is proposed. The developed results have been tested on images from the biomedical field. Applicability of this method has been established in other fields, including those mentioned in the analyzed related research such as forensics, veterinary medicine, museum studies, etc. The first version of the method has been analyzed, and significant algorithm flaws have been identified and corrected, particularly enabling more efficient processing of images with rulers containing two-digit numbers. A new digit recognition model – YOLOv7 – has been trained and integrated into the method, effectively correcting issues with inverted images. The number of images processed by the method has been increased to 90 %. Alternative methods of grouping test segments, which represent the method final step, have been investigated. Experimental data has been processed using various grouping methods, in particular, DBSCAN clustering, median, modified z-score, and interquartile range, and errors compared to manually measured values. In the case of using median and DBSCAN clustering methods, a median error of 4.2-4.4 % was achieved, while in specific configurations of the DBSCAN method, the error was 3.1-3.7 %. A web page with a demonstration version of the method has been developed to attract more data and practical applications in solving real problems. An option to choose one of four grouping methods, their parameters (for DBSCAN), and images for method testing has been provided. The visualization of a random test segment on the uploaded image has been implemented for a more straightforward subjective evaluation of the obtained result by the user. Future proposals include as follows: conduct additional training of digit detection ...